ilotoki_thu/alpaca_lora_4bit
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fix bug

Browse Source
This commit is contained in:
John Smith
2023-03-29 08:09:40 +08:00
parent 1043ded7d9
commit 1719bd0ce3
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
matmul_utils_4bit.py
View File

@@ -106,7 +106,7 @@ def matmul4bit(x, qweight, scales, zeros, groupsize=-1):
if use_new: if use_new:
if auto_switch: if auto_switch:
if np.prod(x.shape[:-1]) > auto_switch_thd: if np.prod(x.shape[:-1]) > auto_switch_thd:
output = _matmul4bit_v1_recons(x, qweight, scales, zeros) output = _matmul4bit_v1_recons(x.to(scales.dtype), qweight, scales, zeros)
else: else:
output = _matmul4bit_v1(x, qweight, scales.float(), zeros.float()) output = _matmul4bit_v1(x, qweight, scales.float(), zeros.float())
else: else:
@@ -116,7 +116,7 @@ def matmul4bit(x, qweight, scales, zeros, groupsize=-1):
if use_new: if use_new:
if auto_switch: if auto_switch:
if np.prod(x.shape[:-1]) > auto_switch_thd: if np.prod(x.shape[:-1]) > auto_switch_thd:
output = _matmul4bit_v2_recons(x, qweight, scales, zeros, groupsize) output = _matmul4bit_v2_recons(x.to(scales.dtype), qweight, scales, zeros, groupsize)
else: else:
output = _matmul4bit_v2(x, qweight, scales.float(), zeros, groupsize) output = _matmul4bit_v2(x, qweight, scales.float(), zeros, groupsize)
else: else: